Financial time series representation, visualization and mining

Abstract

Financial time series has its own characteristics over other time series data. It is typically characterized by a few salient points and multi-resolution consideration is always necessary for long-term and short-term analyses. In addition, technical analysis is usually used to identify patterns of market behavior, which have high probability to repeat themselves. These patterns are similar in the overall shape but with different amplitudes and/or durations. While the traditional time series representation schemes may not be effective for handling these characteristics, there is a need to revise the existing technologies and build new framework and methods to carry out various financial time series data mining tasks. In this research, financial time series is represented according to the importance of data points. With the concept of data point importance, a novel time domain framework for representing time series is developed. A tree data structure is proposed to represent time series and it can be used to access the time series data according to the order of importance. One of the most significant advantages of the proposed framework over the traditional time series data representations is that it provides a mechanism for compressing a time series in different resolutions while the overall shape of the time series can still be preserved. In addition, it facilitates incremental updating and subsequence retrieval of the time series data. Based on the proposed framework and representation, different applications were investigated. First, two technical content-based searching approaches are proposed. Second, the usage of the tree representation in time series segmentation is demonstrated. In addition, an evolutionary segmentation algorithm based on a given set of pattern templates is introduced. Third, multi-resolution time series visualization for mobile finance applications is realized. Moreover, a visual mining tool for visualizing frequently appearing and surprising patterns from time series across different resolutions is developed. Finally, the process of discovering frequently appearing patterns using clustering technique is discussed. Through these applications, encouraging experimental results were obtained and the proposed methods were found particularly effective in financial time series data.